1023 
3D RECONSTRUCTION FOR A CULTURAL HERITAGE VIRTUAL TOUR SYSTEM 
Y. Bastanlar 3 ’*, N. Grammalidis b , X. Zabulis c , E. Yilmaz 3 , Y. Yardimci 3 , G. Triantafyllidis b 
a Informatics Institute, Middle East Technical University, Ankara, Turkey - (yalinb, eyilmaz, yardimy)@ii.metu.edu.tr 
b Informatics and Telematics Institute, CERTH, Thessaloniki, Greece - (ngramm, gatrian)@iti.gr 
c Foundation for Research and Technology - Hellas, Inst, of Computer Science, Heraklion, Greece - 
zabulis@ics.forth.gr 
KEY WORDS: Cultural Heritage, Reconstruction, Web-based, Visualization, GIS, Virtual Reality 
ABSTRACT: 
The aim of this study is to build a Web-based virtual tour system, focused at the presentation of archaeological sites. The proposed 
approach is comprised of powerful techniques such as multiview 3D reconstruction, omnidirectional viewing based on panoramic 
images, and their integration with GIS technologies. In the proposed method, the scene is captured from multiple viewpoints 
utilizing off-the-shelf equipment and its 3D structure is extracted from the acquired images based on stereoscopic techniques. Color 
information is added to the generated 3D model of the scene and the result is converted to a common 3D scene modeling format. 
The 3D models and interactive virtual tour tools such as 360° viewing are integrated with GIS technologies in which the excavation 
site plans can be added as detailed raster overlays. 
1. INTRODUCTION 
Web-based virtual tour applications constructed by 360° 
panoramic images are started to being used extensively all over 
the world. Effectiveness and usability of these tours were 
discussed by Bastanlar (2007) and Villaneuva et al. (2004). 
Usage of image-based 3D reconstructions in virtual tours is 
limited due to their present day lower quality is not appealing 
yet. Cultural heritage is one of the most important application 
areas of these technologies. Example studies (Guamieri et al., 
2004; Kadobayashi et al., 2004; Conforti Andreoni and Pinto, 
2004) on 3D reconstruction of cultural heritage were performed 
by merging image data with the output of 3D laser scanner. 3D 
scanner technology is efficiently developed to scan the 
environment and add color information to generate the 3D 
model. However, the necessary equipment is still very 
expensive and capturing the 3D data and post-processing is very 
time-consuming. 
In this paper, automatic and photorealistic 3D scene 
reconstruction from images is used to create content for a 
cultural heritage virtual tour system. With the same aim, Grim 
et al. (2002) worked to generate 3D reconstruction of a 
demolished Buddha statue. However the 3D model is not 
impressive due to the usage of limited number and low quality 
photographs. Better results were obtained by Pollefeys et al. 
(1999) who used the recordings by a video camera. Later they 
applied their technique for the Sagalassos archaeological site 
(Pollefeys et al., 2004). 
The 3D scene could be synthesized, e.g. by a 3D modeller, by 
performing surface modelling and then adding texture 
information. Some of the current WWW applications are 
composed of graphical textures which are displayed via a 
VRML plug-in. The problem with such synthesized 3D models 
for cultural applications is that the feeling of reality is lost and 
the procedure to generate them is tedious and requires highly- 
experienced personnel. 
An automatic procedure for accurate and photorealistic scene 
modeling that is efficient in terms of the computational 
resources is not straightforward. We propose a pipeline for 
reconstruction and presentation of archaeological sites. In short, 
the steps are: 
1) Acquisition of multiple high-resolution images or video 
recording and subsequent selection of key frames. 
2) Computation of internal camera calibration parameters. 
3) Estimation of lens distortion and image rectification. 
4) Extrinsic calibration of the acquired images, based on 
robust feature extraction, tracking and camera motion 
estimation techniques, 
5) Multi-view stereo reconstruction of the scene using the 
acquired images and intrinsic and extrinsic calibration 
parameters. 
6) Conversion of the reconstruction output to textured VRML 
format, which includes triangulation of points into a mesh, 
combination of textures from different images 
7) Generation of KML/KMZ file from VRML format. 
8) Display of the reconstructed portion of the archeological 
site with the excavation site plans as detailed raster 
overlays, on the Google Earth™ system or other GIS tools 
that support KML/KMZ format. 
As is the case for most multiview stereo reconstruction 
techniques, the accuracy of the final results greatly depends on 
the quality of both camera calibration and motion estimation 
(Steps 2 to 4). To efficiently tackle the problem of fully- 
automatic motion estimation, the proposed approach employs 
state-of-the-art techniques (Beardsley et al., 1997; Pollefeys et 
al., 1999; Pollefeys et al., 2004). Custom modifications were 
made to these techniques to improve accuracy of the calibration 
results, namely, robust feature point detection and matching 
using SIFT (Lowe, 2004) and bundle adjustment (Lourakis and 
Argyros, 2004). Details are presented in Section 2. 
* Corresponding author.